Re: [PATCHv4] shmem: avoid huge pages for small files

[Hugh Dickins <hughd@google.com>]

On Sat, 22 Oct 2016, Kirill A. Shutemov wrote:
> 
> Huge pages are detrimental for small file: they causes noticible
> overhead on both allocation performance and memory footprint.
> 
> This patch aimed to address this issue by avoiding huge pages until file
> grown to size of huge page. This would cover most of the cases where huge
> pages causes regressions in performance.
> 
> Couple notes:
> 
>   - if shmem_enabled is set to 'force', the limit is ignored. We still
>     want to generate as many pages as possible for functional testing.
> 
>   - the limit doesn't affect khugepaged behaviour: it still can collapse
>     pages based on its settings;
> 
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Sorry, but NAK.  I was expecting a patch to tune within_size behaviour.

> ---
>  Documentation/vm/transhuge.txt | 3 +++
>  mm/shmem.c                     | 5 +++++
>  2 files changed, 8 insertions(+)
> 
> diff --git a/Documentation/vm/transhuge.txt b/Documentation/vm/transhuge.txt
> index 2ec6adb5a4ce..d1889c7c8c46 100644
> --- a/Documentation/vm/transhuge.txt
> +++ b/Documentation/vm/transhuge.txt
> @@ -238,6 +238,9 @@ values:
>    - "force":
>      Force the huge option on for all - very useful for testing;
>  
> +To avoid overhead for small files, we don't allocate huge pages for a file
> +until it grows to size of huge pages.
> +
>  == Need of application restart ==
>  
>  The transparent_hugepage/enabled values and tmpfs mount option only affect
> diff --git a/mm/shmem.c b/mm/shmem.c
> index ad7813d73ea7..49618d2d6330 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1692,6 +1692,11 @@ static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
>  				goto alloc_huge;
>  			/* TODO: implement fadvise() hints */
>  			goto alloc_nohuge;
> +		case SHMEM_HUGE_ALWAYS:
> +			i_size = i_size_read(inode);
> +			if (index < HPAGE_PMD_NR && i_size < HPAGE_PMD_SIZE)
> +				goto alloc_nohuge;
> +			break;
>  		}
>  
>  alloc_huge:

So (eliding the SHMEM_HUGE_ADVISE case in between) you now have:

		case SHMEM_HUGE_WITHIN_SIZE:
			off = round_up(index, HPAGE_PMD_NR);
			i_size = round_up(i_size_read(inode), PAGE_SIZE);
			if (i_size >= HPAGE_PMD_SIZE &&
					i_size >> PAGE_SHIFT >= off)
				goto alloc_huge;
			goto alloc_nohuge;
		case SHMEM_HUGE_ALWAYS:
			i_size = i_size_read(inode);
			if (index < HPAGE_PMD_NR && i_size < HPAGE_PMD_SIZE)
				goto alloc_nohuge;
			goto alloc_huge;

I'll concede that those two conditions are not the same; but again you're
messing with huge=always to make it, not always, but conditional on size.

Please, keep huge=always as is: if I copy a 4MiB file into a huge tmpfs,
I got ShmemHugePages 4096 kB before, which is what I wanted.  Whereas
with this change I get only 2048 kB, just like with huge=within_size.

Treating the first extent differently is a hack, and does not respect
that this is a filesystem, on which size is likely to increase.

By all means refine the condition for huge=within_size, and by all means
warn in transhuge.txt that huge=always may tend to waste valuable huge
pages if the filesystem is used for small files without good reason
(but maybe the implementation needs to reclaim those more effectively).

Hugh